From the documentation it seems as if that additionally to erroring when impls are missing or no callback that fits the impl, only @doc false is implied on @impl.
Though, in general, whether you have @impl or not, dialyzer will check you function against the @callback type when you implement a callback.
If the spec is a subset of what is described per the callback, then it doesn’t clash. I’m not sure though how dialyzer looks at it, also, as the interface is defined from the outside world, I’d not try to re(de)fine the spec.
You can provide a typespec for behaviour callback functions. If you do so, and some arg or the return type in your spec is not a subtype of the callback spec, dialyzer will emit a corresponding warning. The same will happen if the types inferred from the callback code do not match spec types.
So e.g. the following GenServer code:
@type state :: %{foo: integer}
@spec handle_call(:foo, GenServer.from(), state) :: {:reply, :ok, state}
def handle_call(:foo, _from, state) do
{:reply, :ok, %{state | bar: 1}}
end
Will lead to a dialyzer warning:
Invalid type specification for function 'Elixir.TestServer':handle_call/3.
The success typing is
('foo', _, #{'bar' := _, _ => _}) ->
{'reply', 'ok', #{'bar' := 1, _ => _}}
which means that either the spec or the impl is wrong.
So far, I didn’t got into a habit of writing specs for callbacks, but I’ve been toying with this idea for some time. I think it would help documenting expected callback args, and might help catching some errors.
I’ve been toying with this idea as well, but I left it because overall in my current environment people feel it’s duplicated documentation that adds more noise and makes the code harder to maintain (that’s why I created this post).
Overall I like the idea of specifying the sub-type of the callback I am about to return for clarity’s sake, but if there is also the chance of causing dialyzer conflicts (as @NobbZ pointed out) then I am ok with leaving it be until I or someone else better understands how dialyzer works inside.